Sign in
VERIFIEDNov 222026

Is Raw Qwen3 4B Instruct 2507 direct logits state of the art?#389Best current rank: #389 in Decision models, as of Nov 22, 2026.

As of Nov 22, 2026, Raw Qwen3 4B Instruct 2507 direct logits is ranked #389 in Decision models (since Nov 22, 2026; sources JevBench; methodology v1.0). Ranked #389 in Decision models.

Computed automaticallyas of Nov 22, 2026methodology v1.0confidence: medium
View as of point-in-time crowns
At a glance

Facts and scores

Raw Qwen3 4B Instruct 2507 direct logits is an open-weights (Apache-2.0) text model from Alibaba / Qwen; 4B parameters and no public price.

How it scores

  • On JevBench price it scores 1.7%, #44 of 134 (top 33%); 1.6 pts behind verdict-small (Manavarya09, multilingual-e5-small 118M); level with Quyet-1.0-Medium (Chinh Nguyen, Qwen3.5-4B decoder, option-letter logits) above and level with open-alternative-jev (Qwen3.5-4B, IkerMoel) below.
  • On JevBench it scores 34.6 (CI 34.1–35.1), #116 of 134 (top 87%); 48.2 behind Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout); Raw Phi-4 mini direct logits is 0.7 above and kev 0.5B is 1.7 below.

History

First result on isitsota: Oct 8, 2026. Best overall rank so far: #389 in Decision models, first held Nov 22, 2026.

$0.0166 per 1,000 decisions.

Receipts

Evidence

Scores come from the latest ingested snapshot
SourceScoreDateTypeReceipt
JevBench
Raw Qwen3 4B Instruct 2507 direct logits
34.6 34.1–35.12026-10-08INDEP.snap_jevbench_2026-10-08_v1.6.1|jevbench-v1.6.1
Reign history

Where it sits in the succession

Full timeline →

Crown intervals for this model are still empty.

Honest about the gaps

Where it is not the crown

CategoryThis modelRankLeader
Decision models#389 in this category#389karpo-decision-31B
Links

Related pages

Is Raw Qwen3 4B Instruct 2507 direct logits SOTA?

Best current rank: #389 in Decision models, as of Nov 22, 2026.

What is Raw Qwen3 4B Instruct 2507 direct logits best at?

As of Nov 22, 2026, Raw Qwen3 4B Instruct 2507 direct logits is ranked #389 in Decision models (since Nov 22, 2026; sources JevBench).